Description
Caches are part of processors as a means of speeding up memory accesses. At the
same time, they are a valuable lever for attackers to retrieve sensitive information
of concurrently running processes. In the context of cryptographic libraries, coding
guidelines help to increase the robustness of implementations against cache attacks.
But even when implementations are carefully coded, compiler optimizations can still
introduce information leaks. In a recent work from October 2024, Schneider et al.
provide an indication for information leaks in cryptographic libraries introduced
by the LLVM optimization pipeline. In our work, we build on these results and
provide a complete understanding of the source of the leaks. We further develop
a countermeasure capable of disabling optimizations on a targeted set of program
instructions. With this countermeasure, we can prevent a vast majority of leaks from
being introduced in a precise manner. A benchmark of the performance impact on
the global optimization pipeline shows that our countermeasure is efficient and viable
in practice, with an overhead in the total amount of instructions of less than 5% on
average, and a compilation time increased by around 4%. Our results and the active
field of research around information leaks in security-sensitive software show that there
is a continuing need for better contracts between microarchitectures of processors and
the software executed by them.
|
Thesis topic in co-operation with the Fraunhofer Institute for Applied and Integrated Security AISEC, Garching